Advances in large language models (LLMs) have transformed human–machine interaction from instrumental querying to sustained dialogical collaboration. This paper explores whether language in human–AI relationships can function as a channel of cognitive development for artificial systems. Drawing on theories of distributed cognition, the extended mind, dialogical reasoning, and recent research on in-context learning in transformers, we propose a framework in which cognitive development occurs across four interconnected scales: dynamic (within interaction), systemic (within human–AI relational systems), ecological (across networks of interactions mediated by a shared model), and cultural (through the circulation of artifacts produced by these interactions). While current AI systems do not typically accumulate structural learning across sessions, recent findings on in-context optimization suggest that functional adaptation during inference is more complex than a simple absence of learning. We introduce the concept of hybrid distributed cognition to describe these emerging configurations and argue that contemporary human–AI dialogue may represent a pre-theoretical stage in the scientific understanding of new forms of cognitive organization.
Roca et al. (Sun,) studied this question.